Financial Sentiment LLM (85% Accuracy)

Metric / Dataset	FinBERT Full	LoRA
Overall Accuracy	85.4%	83.2%
Macro F1-Score	0.83	0.80
PhraseBank	95.9%	97.1%
Twitter	83.3%	80.5%
FiQA	81.5%	72.6%

Live Demo

Model Architecture & Training

This project implements a Multi-Task FinBERT architecture with dual prediction heads:

Component	Purpose
Classification Head	Predicts Negative/Neutral/Positive for news & social media
Regression Head	Predicts continuous sentiment scores for forum discussions

Why Multi-Task FinBERT: Comparing Multi-Task vs. Single-Task FinBERT on the same data revealed a +4.4% accuracy boost. The improvement was especially significant on Twitter (+6.1%), where capturing sentiment nuance is critical.

We benchmarked multiple models (BERT-Base, DistilBERT, FinBERT) and training approaches (Full Fine-Tuning, LoRA):

FinBERT (Full): Best overall accuracy (85.4%) and most consistent behaviour across domains.
LoRA: 99% smaller storage footprint (5MB vs 420MB) but struggles with forum (FiQA) discussions.

Training Data

Dataset	Type	Purpose
Financial PhraseBank	Professional News	Formal financial language
Twitter Financial News	Social Media	Short-form, informal commentary
FiQA Sentiment	Forum Discussions	Regression target for sentiment intensity

Project Roadmap

This project is being developed in three phases:

✅ Phase 1: Baseline - Established single-task benchmarks (Completed Dec 2025).
✅ Phase 2: Optimisation - Implemented Multi-Task architecture and LoRA (Completed Jan 22, 2026).
🔄 Phase 3: Integration (Current) - Deploying sentiment scores as features into the Financial-ML Equity Selection System.

Quick Start

The project is built with PyTorch and HuggingFace Transformers. It includes a custom CLI for training and evaluation.

# Clone and setup
git clone https://github.com/pmatorras/financial-sentiment-llm.git
cd financial-sentiment-llm
python -m venv .venv
source .venv/bin/activate
pip install -e .

# Train the multi-task model
python -m finsentiment train

# Evaluate performance
python -m finsentiment Evaluate

# Run the demo locally
pip install -e ".[demo]"
python app.py  # Opens at localhost:7860

For advanced configuration and flags, see the GitHub Repository.